Implement 3D Gaussian splatting viewer

[ssheorey]

• I/O
• Viewing

Milestones:

• Build Open3D on CPU and review t point cloud I/O |
• Read 3DGS ply files into t point cloud and visualize with view-independent rendering -> PR 1 |
• Write t point cloud into 3DGS ply file -> PR 2 with unit test for read-write-read cycle check. |
• Study 3DGS implementations and identify best option, risks and blockers for integration |
• Study filament renderer and Open3D-filament integration |
• 3DGS Rendering implementation -> PR 3|
• Performance optimization -> PR 4
• Demos -> PR 5

Resources:

Pre-trained INRIA models: https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/datasets/pretrained/models.zip
Optional Validate display of 3DGS .ply / .splat files from luma.ai and polycam
Vulkan compute shaders implementation: https://github.com/shg8/3DGS.cpp/tree/main [LGPL-v2.1]
WebGL implementation using vertex/fragment shader: https://github.com/kishimisu/Gaussian-Splatting-WebGL

Components: (vkGS -> O3D GS)

• GLM
• Vulkan Memory Allocator
• VkRadixSort

Technical open questions:

• How will integration with Filament work? (e.g. Vulkan callbacks, shared Vulkan buffers).
• An alternative is to use Vulkan directly, and only give the final rendered image to filament to display.
• For the initial implementation, round trip to system RAM is OK, but after optimization we want all data and operations on the GPU.